Exploiting Backdoor Trigger Towards Unlearnable Examples

Abstract

Unlearnable examples (UEs) aim to protect personal privacy from abuse in model training. Existing work primarily generates UEs by crafting imperceptible perturbations through search or optimization and introducing them to clean samples. Despite considerable progress in UE methods, their applicability in real-world scenarios and effectiveness under different defense strategies remains limited. In this paper, we reveal the connections between poisoning-based backdoor attacks and UEs from the perspective of shortcut learning. Specifically, both the perturbations in UEs and backdoor triggers are easy-to-learn features, and thus models trained on samples with them will ignore other important semantic features and perform poorly on clean data. Motivated by the observation, we propose TRIGGERUE, which utilizes the backdoor triggers to craft effective UEs for the first time. Our experimental results show that our method significantly outperforms existing state-of-the-art UE methods. Code: https://github.com/pppppkun/TriggerUE.